AgenticTwin wires an LLM reasoning layer onto digital-twin anomaly detection and tests whether small open models can hold it

arXiv·low signalXBlueskyLinkedInCopy link

AgenticTwin (submitted 2026-08-12) grounds LLM-generated explanations in the output of a digital-twin-based anomaly classifier so operators of cyber-physical systems can ask natural-language questions about detected anomalies instead of reading raw sensor volume. Alongside the framework the authors release a benchmark-oriented evaluation pipeline built on synthetic anomalies injected into a real-world weather sensor dataset, enabling controlled generation of operator queries over anomaly events. They specifically evaluate lightweight open-source LLMs for on-site deployment, finding structured agent collaboration plus knowledge-grounded reasoning is what carries performance — relevant to anyone putting agents on constrained industrial hardware.